StreamDecoder#

class scikitplot.cleanprompt.StreamDecoder(guard)[source]#

Decode a model’s reply as it streams in, never splitting a label.

Parameters:

guardGuard

Supplies the vault and the placeholder grammar.

Parameters:

guard (Guard)

Notes

User notes. Call feed with each chunk and print what it returns; call flush once at the end. The text you print is exactly what decoding the whole reply at once would give.

Developer notes. Two things can be cut by a chunk boundary: a bracket label ([EMA | IL-1]) and a literal stand-in (Marion | Holt, or id_1 | 0 where id_10 is also a stand-in). The decoder emits everything before the earliest point where either could still be in progress, and keeps the rest:

• from the last unclosed opening delimiter within MAX_LABEL characters, including a backslash escaping it;

• the longest tail of the buffer that is a proper prefix of a literal stand-in, or a whole stand-in that is also a prefix of a longer one.

Everything emitted is decoded with the same function as a whole reply.

One reply, one audit event. A reply decoded whole records one decoded event. Streamed, it arrives in as many chunks as the model emits tokens, and recording each chunk turned one reply into thousands of events that said nothing a single one does not (CP-090). The counts are gathered as chunks arrive and recorded once, by flush, with the number of chunks; they equal the counts of the same reply decoded whole. A stream that is abandoned without flush records no event; flush is what ends a reply.

feed(chunk)[source]#

Add a chunk and return the decoded text that is now certain.

Parameters:

chunkstr

The next piece of the reply.

Returns:

str

Decoded text; possibly empty while a label is incomplete.

Raises:

TypeError

If chunk is not a string.

Parameters:

chunk (str)

Return type:

str

flush()[source]#

Return the decoded remainder, record the reply, and reset.

Returns:

str

Whatever was held back, decoded.

Return type:

str